The present invention relates generally to machine diagnostics, and more specifically, to a system and method for processing historical repair data and operational parameter data for predicting one or more repairs from new operational parameter data from a malfunctioning machine.
A machine such as locomotive includes elaborate controls and sensors that generate faults when anomalous operating conditions of the locomotive are encountered. Typically, a field engineer will look at a fault log and determine whether a repair is necessary.
Approaches like neural networks, decision trees, etc., have been employed to learn over input data to provide prediction, classification, and function approximation capabilities in the context of diagnostics. Often, such approaches have required structured and relatively static and complete input data sets for learning, and have produced models that resist real-world interpretation.
Another approach, Case Based Reasoning (CBR), is based on the observation that experiential knowledge (memory of past experiences—or cases) is applicable to problem solving as learning rules or behaviors. CBR relies on relatively little pre-processing of raw knowledge, focusing instead on indexing, retrieval, reuse, and archival of cases. In the diagnostic context, a case generally refers to a problem/solution description pair that represents a diagnosis of a problem and an appropriate repair. More particularly, a case is a collection of fault log and corresponding operational and snapshot data patterns and other parameters and indicators associated with one specific repair event in the machine under consideration.
CBR assumes cases described by a fixed, known number of descriptive attributes. Conventional CBR systems assume a corpus of fully valid or “gold standard” cases that new incoming cases can be matched against.
U.S. Pat. No. 5,463,768 discloses an approach which uses error log data and assumes predefined cases with each case associating an input error log to a verified, unique diagnosis of a problem. In particular, a plurality of historical error logs are grouped into case sets of common malfunctions. From the group of case sets, common patterns, i.e., consecutive rows or strings of data, are labeled as a block. Blocks are used to characterize fault contribution for new error logs that are received in a diagnostic unit.
For a continuous fault code stream where any or all possible fault codes may occur from zero to any finite number of times and the fault codes may occur in any order, predefining the structure of a case is nearly impossible.
U.S. Pat. No. 6,343,236, assigned to the same assignee of the present invention, discloses a system and method for processing historical repair data and fault log data, which is not restricted to sequential occurrences of fault log entries and which provides weighted repair and distinct fault cluster combinations, to facilitate analysis of new fault log data from a malfunctioning machine. Further, U.S. Pat. No. 6,415,395, assigned to the same assignee of the present invention, discloses a system and method for analyzing new fault log data from a malfunctioning machine in which the system and method are not restricted to sequential occurrences of fault log entries, and wherein the system and method predict one or more repair actions using predetermined weighted repair and distinct fault cluster combinations. Additionally, U.S. Pat. No. 6,336,065, assigned to the same assignee of the present invention, discloses a system and method that uses snapshot observations of operational parameters from the machine in combination with the fault log data in order to further enhance the predictive accuracy of the diagnostic algorithms used therein.
It is believed that the inventions disclosed in the foregoing patent applications provide substantial advantages and advancements in the art of diagnostics. It would be desirable, however, to provide a system and method that uses anomaly definitions based on operational parameters to generate diagnostics and repair data. The anomaly definitions are different from faults in the sense that the information used can be taken in a relatively wide time window, whereas faults, or even fault data combined with snapshot data, are based on discrete behavior occurring at one instance in time. The anomaly definitions, however, may be advantageously analogized to virtual faults and thus such anomaly definitions can be learned using the same diagnostics algorithms that can be used for processing fault log data.